Alibaba's OpenGCUA cuts agent steps by 53% with hybrid runtime

Alibaba released OpenGCUA, an open framework for benchmarking and running generalist computer-use agents across OSWorld-v2 and real macOS desktops. A hybrid runtime combining Python code execution with on-demand screen observation cut required steps per task by 53% and raised mean partial-completion scores by 19.6 points across eight models, including GPT-6 Astra, Claude Opus 5, Gemini 3.8 Flash and Qwen3.8-Max.

• OpenGCUA benchmarks agents on OSWorld-v2 and real macOS desktops
• Hybrid runtime cut steps per task by 53% across eight models
• Mean partial-completion scores rose 19.6 points
• Qwen3.8-Max used 19.4x more reasoning tokens than GPT-6 Astra